Hi team, I have one question regarding: why recurrent model (AutoLSTM,...) only supports step size = 1 when we do evaluation (time series cross validation). Wondering if this is a theoretical problem or implementation one. Thank you

The LSTM is an autoregressive model, meaning it predicts one step at a time until the entire forecast horizon is covered. That's why step size is always 1.

Thanks for your response Marco. Let me clarify a little: when we used .cross_validation() with a neuralforecast object, parameter "step_size = n" to simply move the forecast by n points ahead and model still makes "h" predictions. I haven't fully get your explanation in this case. I would also love to take this chance to ask you regarding automatic deep learning, such as AutoLSTM, when I fit a df with exogenous variables, are they by default included in the training. The reason for my question is that with base model, I see the tutorials showing that we need to specify historical exog variables. Thank you so much Marco for bearing with me ^^

step_size

controls the distance between each window. When you use recurrent models, it has to be set to 1. Yes, exogenous variables are included by default. Make sure to also have their future values so the predictions work correctly!